A Survey on the Robustness of Feature Importance and Counterfactual\n Explanations

Abstract

There exist several methods that aim to address the crucial task of\nunderstanding the behaviour of AI/ML models. Arguably, the most popular among\nthem are local explanations that focus on investigating model behaviour for\nindividual instances. Several methods have been proposed for local analysis,\nbut relatively lesser effort has gone into understanding if the explanations\nare robust and accurately reflect the behaviour of underlying models. In this\nwork, we present a survey of the works that analysed the robustness of two\nclasses of local explanations (feature importance and counterfactual\nexplanations) that are popularly used in analysing AI/ML models in finance. The\nsurvey aims to unify existing definitions of robustness, introduces a taxonomy\nto classify different robustness approaches, and discusses some interesting\nresults. Finally, the survey introduces some pointers about extending current\nrobustness analysis approaches so as to identify reliable explainability\nmethods.\n